Fixed test scenario. Updated ra with optimization discussed yesterday. Tests pass, inferrence is slow on the 5th model.

Some updates. We finally have some learner level unittests. Separated inference from model export, so we can test this feature in isolation.